Distributing Responsibility: A Learning Method That Complies with Dale's Principle Without Weight Transport

Real neurons follow Dale's principle, mainly having either excitatory or inhibitory properties, but existing artificial neural networks have a limitation in that all units mix positive and negative weights. Also, Backpropagation carries the Weight Transport problem, in which the transpose of the feedforward weight matrix is required during backpropagation.

This study extends the Error Diffusion (ED) approach to present a method that can learn while strictly complying with Dale's principle, without weight transport. Each layer is separated into distinct excitatory and inhibitory streams, and Modulo error routing is designed to work effectively even in multi-class problems.

This mechanism was also shown to be valid in PPO training within reinforcement learning environments. When tested in the Ant, Humanoid, and HalfCheetah environments, the proposed Modulo routing approach showed the highest agreement with actual Backpropagation gradients. This demonstrates the potential for learning rules that are both more biologically plausible and efficient.